Heatmap

# A tibble: 2 × 3
  class    precision recall
  <fct>        <dbl>  <dbl>
1 City Pop     0.774  0.774
2 U.S. Pop     0.774  0.774

In order to compute the model, I did capped the playlists at 31 songs in each group. The accuracy of the model as of right now is by the use of the formula (TP+TN)/total according to this website is (25+23)/62 = 0.77.., thus according to various sources like this one it is a relatively good accuracy while at the same time being within a realistic interval.

By this, I can assume that the model is actually very good at classifying exactly what City Pop and what U.S Pop is. In the next header, I will see what kind of labels were the most important in the classification of these playlists, and whether my visualizations from the earlier weeks were any sort of effective.

None of the features I used in visualizations were effective!

# A tibble: 2 × 3
  class    precision recall
  <fct>        <dbl>  <dbl>
1 City Pop     0.84   0.677
2 U.S. Pop     0.730  0.871

          Truth
Prediction City Pop U.S. Pop
  City Pop       21        4
  U.S. Pop       10       27

I hope this is the right way of getting the confusion matrix out of the random forest model. If I use the same formula as in the previous slide, i will get an accuracy of 0.77…, just as the last one.

Regarding the feature selection, it is shown in the graph that the timbre vector c11, followed by the track level feature loudness, followed yet again by timbre vector c1. This is quite humorous as it was said in one of the lectures that this timbre vector is the rough equivalent to loudness. After those, primarily timbre features are the ones that are of the most importance. This is quite interesting, as I did mention at the very start of the course (without any knowledge of the terms of music) said that City Pop has a lot of layers to them. This was meant as “there is a lot going on” and it sounds different.

For the final portfolio, I will use these features in order to improve the existing visualizations I have (and I have the ability to remove some of them). Yay!

Histogram of tempi

Here is the distribution of tempi within my corpus, divided between the Japanese group and the U.S. group. As the graphs show, there are distinct differences in which tempo is preferred in the western world - which one can see is around 120 bpm. Interestingly, this corresponds well to the research that has been conducted where 120-125 bpm seems to be a natural tempo for humans and a link between the natural tempo and bpm in music has been found. However, according to the corpus I am using, this is not the case for the eastern world. The distribution here is more even, with more songs being in the range of 100-135 bpm. However, there are more songs that lie around the 105 bpm mark.

Aside from the obvious archetypes in this distribution, there are two outliers from each group. These are the songs “A HOPE FROM SAD STREET” by Anri, and “Love Is a Contact Sport” by Whitney Houston. They each have tempi that lie around 180 and 175 bpm respectively.

Billie Jean - our atypical song

Here is the tempogram of the indentified outlier from the graph of energy’s effect on danceability, where the energy is low, but danceability is high. This is a completely standard tempogram, in terms of the overall execution of the formulations on this piece as there are no disturbances or anything that went wrong. “Billie Jean” is a very stereotypical song in this regard, with a bpm of approximately 120 and very steady beat, rhythm and tempo that do not seem to change. This is probably a part of the reason the overall danceability of this piece is very high, in spite that its energy is very low.

A Hope From Sad Street - An Outlier in Distribution

Providing the distribution of tempi in the form of a histogram, there was a couple outliers in both genres. “A HOPE FROM SAD STREET” by Anri is one of them. Here, in comparison to “Billie Jean” there were apparantly some issues when generating the tempo, as there are multiple yellow lines that flicker across the entire piece. However, as one can tell, it was not a complete failure, as it was able to generate not one, but two tempo octaves. It is not as stable as the previous piece, but the lines are clearly there.

This flickering could be attributed to the way this song is realized, in that there are many layers of sound, there are the bassline, but there are also a layer of trumpet as well as a choir and points where the song is stopped and picked up again. The bassline often is replaced by piano, and guitar solos - before returning to the “status quo”. As far as I can tell, these instances are represented in the points in the tempogram.

Love Is a Contact Sport - Western Outlier in Distribution

In comparison to Anri’s song, “Love is a Contact Sport” by Whitney Houston does not have any issues at all, as clearly shown by the image. The song is also realized very clearly, with a clear tempo throughout the entire song. This song was also one of the outliers in the distribution chart, and was the western counterpart to Anri’s. Similarly to the aforementioned song, this has two tempo-octaves.

Interesting find - Cat’s Eye by Anri

Once again, here is a song by Anri, called “CAT’S EYE - (NEW TAKE)”. I wanted to include this, due to the algorithm being able to generate the tempo very clearly in the beginning but when the song has reached around 130 seconds, there is a clear window where it was not able to generate the tempo. This is very clear if you listen to the song, as there is a moment around that time where the bass guitar has a solo. To me, it feels quicker than the song itself, which it has been able to show in some regard - however, this shift was clearly something the tempogram could not handle and was interesting to me.

Histogram of keys

As one can see from the graph here, the distribution of keys between each song in each of my corpus group are, overall, quite evenly distributed. However, one can tell that different groups often prefer different keys. City Pop seems to prefer keys that are in C, D, F, and G for the most part, whereas U.S. Pop, while sharing a decent amount of songs in C as well, seem to also prefer D#, A, and B - unlike City Pop. It should be noted however, that there are a few songs more in the city pop group than the U.S. one, so it can skew the distributions.

Mayonaka no Door/Stay With Me City Pop

4:00A.M. City Pop

Remember Summer Days City Pop

I Wanna Dance With Somebody (Who Loves Me) Pop

Material Girl Pop

Thriller Pop

Distribution of energy between City Pop and US Pop

Here are plots to give insight in the general tempos effect on energy in the groups City Pop and US Pop. In the first histogram plot, one can see that the distribution of energy is more even for pop than for city pop. It says that it counted 15 of the data in my playlist for City Pop to have an energy of a little over 0.6.

Tempo and Energy UPDATED

Danceability on Valence UPDATED

Energy on danceability NEW

Here is a plot of the effect of energy on danceability, with size of the plots as well as the band around the line indicating the tempo of the songs. One can see that there is more of a linear trend with US Pop, indicating that up until around energy of 0.5 there seems to be a correlation between energy and danceability. Tempo, however, do seem to have no pattern at first glance. Whereas for the City Pop playlist, there seem to be a slight curve in the beginning of the graph, but overall there is a very even trend of the effect of energy on danceability. At first glance, there also seem to be no indication that there is a trend for tempo.

However, once can tell that U.S. Pop seem to have more of a positive linear correlation between energy and danceability - more than its eastern counterpart in any case.

In both groups, there are two outliers that especially draw one’s eye - which is “Billie Jean” by Michael Jackson, and “横顔” by Taeko Onuki.